A Multi-level Causal Inference Method for Population Downscaling(2000)

This study develops a new approach, multi-level causal Bayesian network spatialization (MCBNS), which integrates causal inference with multi-level modeling. This method employs structural causal models to analyze direct and indirect relationships between population distribution and influencing factors at the county level, establishing stable causal relationships. It then refines these relationships using data from selected townships to build multi-level models for more accurate prediction. Experimental results show that this strategy not only outperforms single causal models or random forests, but it also achieves validation accuracy exceeding existing product datasets (WorldPop and LandScan) by more than 10%. Notably, accuracy gains are more pronounced in non-urban areas. With MCBNS, population distributions at one square kilometer grids of China in 2000, 2010, and 2020 were produced.